public void AddFrom(Vol V)
{
for (int k = 0; k < this.W.Length; k++)
{
this.W[k] += V.W[k];
}
}
public Vol Forward(Vol V, bool is_training)
{
is_training = false;
this.input = V;
var V2 = V.Clone();
var N = V.W.Length;
if (is_training)
{
// do dropout
for (var i = 0; i < N; i++)
{
if ((float)r.NextDouble() < this.drop_prob)
{
V2.W[i] = 0; this.dropped[i] = true;
} // drop!
else
{
this.dropped[i] = false;
}
}
}
else
{
// scale the activations during prediction
for (var i = 0; i < N; i++)
{
V2.W[i] *= 0.5f;
}
}
this.Output = V2;
return(this.Output); // dummy identity function for now
}
public static Vol Image_To_Vol(double[] img, int width, int hight, bool convert_grayscale)
{
var p = img;
var w = width;
var h = hight;
double[] pv = new double[img.Length];
for (var i = 0; i < p.Length; i++)
{
pv[i] = (p[i] / 255.0 - 0.5); // normalize image pixels to [-0.5, 0.5]
}
var x = new Vol(w, h, 4, 0.0);
x.W = pv;
if (convert_grayscale)
{
var x1 = new Vol(w, h, 1, 0.0);
for (int i = 0; i < w; i++)
{
for (int j = 0; j < h; j++)
{
x1.Set(i, j, 0, x.Get(i, j, 0));
}
}
x = x1;
}
return(x);
}
public void AddFromScaled(Vol V, double a)
{
for (int k = 0; k < this.W.Length; k++)
{
this.W[k] += a * V.W[k];
}
}
private void DLRNforWidth(Vol V, double n2, int x)
{
for (var y = 0; y < V.SY; y++)
{
for (var i = 0; i < V.Depth; i++)
{
var chain_grad = this.Output.Get_Grad(x, y, i);
var S = this.S_cache_.Get(x, y, i);
var SB = Math.Pow(S, this.beta);
var SB2 = SB * SB;
// normalize in a window of size n
for (var j = Math.Max(0, i - n2); j <= Math.Min(i + n2, V.Depth - 1); j++)
{
var aj = V.Get(x, y, (int)j);
var g = -aj *this.beta *Math.Pow(S, this.beta - 1) * this.alpha / this.n * 2 * aj;
if (j == i)
{
g += SB;
}
g /= SB2;
g *= chain_grad;
V.Add_Grad(x, y, (int)j, g);
}
}
}
}
public Vol Forward(Vol V, bool is_training)
{
this.input = V;
var A = new Vol(1, 1, this.OutputDepth, 0.0f);
this.Output = V;
return(this.Output);
}
private void LRN(Vol V, Vol A, float n2)
{
var source = Enumerable.Range(0, V.Width);
var pquery = from num in source.AsParallel()
select num;
pquery.ForAll((x) => LRNforWidth(V, A, n2, x));
}
private void ConvFilter(Vol V, Vol tempOutput, int inputWidth, int inputHeight, int xy_stride, int d)
{
var source = Enumerable.Range(0, this.OutputHeight);
var pquery = from num in source.AsParallel()
select num;
pquery.ForAll((ay) => ConvOverRows(V, tempOutput, inputWidth, inputHeight, xy_stride, d, ay));
}
private void DLRN(Vol V, double n2)
{
var source = Enumerable.Range(0, V.SX);
var pquery = from num in source.AsParallel()
select num;
pquery.ForAll((x) => DLRNforWidth(V, n2, x));
}
public static Vol Augment(Vol V, int corp, int dx, int dy, bool fliplr)
{
//if (dx==0)
//{
// dx = Util.Randi(0, V.SX - corp);
//}
//if (dy == 0)
//{
// dy = Util.Randi(0, V.SY- corp);
//}
Vol W;
if (corp != V.SX || dx != 0 || dy != 0)
{
W = new Vol(corp, corp, V.Depth, 0.0);
for (int x = 0; x < corp; x++)
{
for (int y = 0; y < corp; y++)
{
if (x + dx < 0 || x + dx >= V.SX || y + dy < 0 || y + dy >= V.SY)
{
continue;
}
for (int d = 0; d < V.Depth; d++)
{
W.Set(x, y, d, V.Get(x + dx, y + dy, d));
}
}
}
}
else
{
W = V;
}
if (fliplr)
{
var W2 = W.CloneAndZero();
for (int x = 0; x < W.SX; x++)
{
for (int y = 0; y < W.SY; y++)
{
for (int d = 0; d < W.Depth; d++)
{
W2.Set(x, y, d, W.Get(W.SX - x - 1, y, d));
}
}
}
W = W2;
}
return(W);
}
public Vol Clone()
{
var v = new Vol(this.SX, this.SY, this.Depth, 0.0);
var n = this.W.Length;
for (int i = 0; i < n; i++)
{
v.W[i] = this.W[i];
}
return(v);
}
public Vol Clone()
{
var v = new Vol(this.Width, this.Height, this.Depth, 0.0f);
var n = this.W.Length;
for (int i = 0; i < n; i++)
{
v.W[i] = this.W[i];
}
return(v);
}
public Vol Forward(Vol V, bool is_training)
{
var act = this.Layers[0].Forward(V, is_training);
for (int i = 1; i < this.Layers.Count; i++)
{
act = this.Layers[i].Forward(act, is_training);
}
act = this.LossLayer.Forward(act, is_training);
return(act);
}
public Vol Forward(Vol V, bool is_training)
{
this.input = V;
var A = new Vol(this.OutputWidth, this.OutputHeight, this.OutputDepth, 0.0f);
Conv(V, is_training, A);
this.Output = A;
return(this.Output);
}
public static Vol Augment(Vol V, int corp)
{
int dx; int dy; bool fliplr = false;
dx = (int)Util.Randi(0, V.Width - corp);
dy = (int)Util.Randi(0, V.Height - corp);
Vol W;
if (corp != V.Width || dx != 0 || dy != 0)
{
W = new Vol(corp, corp, V.Depth, 0.0f);
for (int x = 0; x < corp; x++)
{
for (int y = 0; y < corp; y++)
{
if (x + dx < 0 || x + dx >= V.Width || y + dy < 0 || y + dy >= V.Height)
{
continue;
}
for (int d = 0; d < V.Depth; d++)
{
W.Set(x, y, d, V.Get(x + dx, y + dy, d));
}
}
}
}
else
{
W = V;
}
if (fliplr)
{
var W2 = W.CloneAndZero();
for (int x = 0; x < W.Width; x++)
{
for (int y = 0; y < W.Height; y++)
{
for (int d = 0; d < W.Depth; d++)
{
W2.Set(x, y, d, W.Get(W.Width - x - 1, y, d));
}
}
}
W = W2;
}
return(W);
}
public Vol Forward(Vol V, bool is_training)
{
this.Input = V;
var tempOutput = new Vol(this.OutputWidth | 0, this.OutputHeight | 0, this.OutputDepth | 0, 0.0);
var inputWidth = V.SX | 0;
var inputHeight = V.SY | 0;
var xy_stride = this.Stride | 0;
for (int d = 0; d < this.OutputDepth; d++)//for each output depth aka filters
{
var f = this.Filters[d];
var x = -this.Pad | 0;
var y = -this.Pad | 0;
for (int ay = 0; ay < this.OutputHeight; y += xy_stride, ay++)// for each out height
{
x = -this.Pad | 0;
for (var ax = 0; ax < this.OutputWidth; x += xy_stride, ax++) // for each out width
{
// xy_stride
// convolve centered at this particular location
var a = 0.0;
for (var fy = 0; fy < f.SY; fy++) // for each element in the filter height
{
var oy = y + fy; // coordinates in the original input array coordinates
for (var fx = 0; fx < f.SX; fx++) // for each element in the filter width
{
//x is current width element of the output
//fx is the current width element of the filter
var ox = x + fx;
if (oy >= 0 && oy < inputHeight && ox >= 0 && ox < inputWidth)
{
for (var fd = 0; fd < f.Depth; fd++) // for each filter depth or input depth
{
// multiply filter pixel by image pixel and add (shared weight filter)
// avoid function call overhead (x2) for efficiency, compromise modularity :(
//filter (fx,fy,fd) *
//input (ox,oy,fd)
a += f.W[((f.SX * fy) + fx) * f.Depth + fd] *
V.W[((inputWidth * oy) + ox) * V.Depth + fd];
}
}
}
}
a += this.Biases.W[d];
tempOutput.Set(ax, ay, d, a);
}
}
}
this.Output = tempOutput;
return(this.Output);
}
public Vol Forward(Vol V, bool is_training)
{
this.in_Act = V;
var A = V.CloneAndZero();
this.S_cache_ = V.CloneAndZero();
var n2 = Math.Floor(this.n / 2);
LRN(V, A, n2);
this.Output = A;
return(this.Output); // dummy identity function for now
}
public Vol Forward(Vol V, bool is_training)
{
this.in_Act = V;
var A = new Vol(this.OutputWidth, this.OutputHeight, this.OutputDepth, 0.0);
var n = 0; // a counter for switches
for (var d = 0; d < this.OutputDepth; d++)
{
var x = -this.Pad;
var y = -this.Pad;
for (var ax = 0; ax < this.OutputWidth; x += this.Stride, ax++)
{
y = -this.Pad;
for (var ay = 0; ay < this.OutputHeight; y += this.Stride, ay++)
{
// convolve centered at this particular location
double a = -99999; // hopefully small enough ;\
var winx = -1; var winy = -1;
for (var fx = 0; fx < this.SX; fx++)
{
for (var fy = 0; fy < this.SY; fy++)
{
var oy = y + fy;
var ox = x + fx;
if (oy >= 0 && oy < V.SY && ox >= 0 && ox < V.SX)
{
var v = V.Get(ox, oy, d);
// perform max pooling and store pointers to where
// the max came from. This will speed up backprop
// and can help make nice visualizations in future
if (v > a)
{
a = v; winx = ox; winy = oy;
}
}
}
}
this.Switchx[n] = winx;
this.Switchy[n] = winy;
n++;
A.Set(ax, ay, d, a);
}
}
}
this.Output = A;
return(this.Output);
}
public Vol Forward(Vol V, bool is_training)
{
this.in_Act = V;
var V2 = V.CloneAndZero();
var N = V.W.Length;
var V2w = V2.W;
var Vw = V.W;
for (var i = 0; i < N; i++)
{
V2w[i] = 1.0 / (1.0 + Math.Exp(-Vw[i]));
}
this.Output = V2;
return(this.Output);
}
public Vol Forward(Vol V, bool is_training)
{
var act = this.Layers[0].Forward(V, is_training);
for (int i = 1; i < this.Layers.Count; i++)
{
act = this.Layers[i].Forward(act, is_training);
if (this.ForwardLayer != null)
{
this.ForwardLayer(this.Layers[i], new EventArgs());
}
}
act = this.LossLayer.Forward(act, is_training);
return(act);
}
public Vol Forward(Vol V, bool is_training)
{
this.input = V;
var V2 = V.CloneAndZero();
var N = V.W.Length;
var V2w = V2.W;
var Vw = V.W;
for (var i = 0; i < N; i++)
{
V2w[i] = (float)Math.Tanh(Vw[i]);
}
this.Output = V2;
return(this.Output);
}
private void Conv(Vol V, bool is_training, Vol A)
{
var n = 0; // a counter for switches
for (int d = 0; d < this.OutputDepth; d++)
{
for (var ax = 0; ax < this.OutputWidth; ax++)
{
var x = -this.Pad;
x += (this.Stride * ax);
for (var ay = 0; ay < this.OutputHeight; ay++)
{
var y = -this.Pad;
y += (this.Stride * ay);
// convolve centered at this particular location
float a = -99999; // hopefully small enough ;\
var winx = -1; var winy = -1;
for (var fx = 0; fx < this.KernelWidth; fx++)
{
for (var fy = 0; fy < this.KernelHeight; fy++)
{
var oy = y + fy;
var ox = x + fx;
if (oy >= 0 && oy < V.Height && ox >= 0 && ox < V.Width)
{
var v = V.Get(ox, oy, d);
// perform max pooling and store pointers to where
// the max came from. This will speed up backprop
// and can help make nice visualizations in future
if (v > a)
{
a = v; winx = ox; winy = oy;
}
}
}
}
this.Switchx[n] = winx;
this.Switchy[n] = winy;
n++;
A.Set(ax, ay, d, a);
}
}
}
//var source = Enumerable.Range(0, this.OutputDepth);
//var pquery = from num in source.AsParallel()
// select num;
// pquery.ForAll((d) => );
}
public Vol Forward(Vol V, bool is_training)
{
this.Input = V;
var tempOutput = new Vol(this.OutputWidth | 0, this.OutputHeight | 0, this.OutputDepth | 0, 0.0f);
var inputWidth = V.Width | 0;
var inputHeight = V.Height | 0;
var xy_stride = this.Stride | 0;
for (int d = 0; d < this.OutputDepth; d++)
{
ConvFilter(V, tempOutput, inputWidth, inputHeight, xy_stride, d);
}
this.Output = tempOutput;
return(this.Output);
}
public Vol Forward(Vol V, bool is_training)
{
this.in_Act = V;
var V2 = V.Clone();
var N = V.W.Length;
var V2w = V2.W;
for (var i = 0; i < N; i++)
{
if (V2w[i] < 0)
{
V2w[i] = 0; // threshold at 0
}
}
this.Output = V2;
return(this.Output);
}
public Vol Forward(Vol V, bool is_training)
{
this.input = V;
var A = new Vol(1, 1, this.OutputDepth, 0.0f);
var Vw = V.W;
for (int i = 0; i < this.OutputDepth; i++)
{
var a = 0.0f;
var wi = this.Filters[i].W;
for (var d = 0; d < this.num_inputs; d++)
{
a += Vw[d] * wi[d];
}
a += this.Biases.W[i];
A.W[i] = a;
}
this.Output = A;
return(this.Output);
}
public Vol Forward(Vol V, bool is_training)
{
this.in_Act = V;
var A = new Vol(1, 1, this.OutputDepth, 0.0);
var Vw = V.W;
for (var i = 0; i < this.OutputDepth; i++)
{
var a = 0.0;
var wi = this.Filters[i].W;
for (var d = 0; d < this.num_inputs; d++)
{
a += Vw[d] * wi[d]; // for efficiency use Vols directly for now
}
a += this.Biases.W[i];
A.W[i] = a;
}
this.Output = A;
return(this.Output);
}
public Vol Forward(Vol V, bool is_training)
{
this.in_Act = V;
var A = new Vol(1, 1, this.Out_Depth, 0.0);
// compute max activation
var ass = V.W;
var amax = V.W[0];
for (var i = 1; i < this.Out_Depth; i++)
{
if (ass[i] > amax)
{
amax = ass[i];
}
}
// compute exponentials (carefully to not blow up)
var es = new double[this.Out_Depth];
var esum = 0.0;
for (var i = 0; i < this.Out_Depth; i++)
{
var e = Math.Exp(ass[i] - amax);
esum += e;
es[i] = e;
}
// normalize and output to sum to one
for (var i = 0; i < this.Out_Depth; i++)
{
es[i] /= esum;
A.W[i] = es[i];
}
this.es = es; // save these for backprop
this.Output = A;
return(this.Output);
}
public Policy policy(double[] s)
{
// compute the value of doing any action in this state
// and return the argmax action and its value
var svol = new Vol(1, 1, this.net_inputs);
svol.W = s;
var action_values = this.value_net.Forward(svol, false);
var maxk = 0;
var maxval = action_values.W[0];
for (var k = 1; k < this.num_actions; k++)
{
if (action_values.W[k] > maxval)
{
maxk = k; maxval = action_values.W[k];
}
}
return(new Policy()
{
action = maxk, value = maxval
});
}
private void LRNforWidth(Vol V, Vol A, double n2, int x)
{
for (var y = 0; y < V.SY; y++)
{
for (var i = 0; i < V.Depth; i++)
{
var ai = V.Get(x, y, i);
// normalize in a window of size n
var den = 0.0;
for (var j = Math.Max(0, i - n2); j <= Math.Min(i + n2, V.Depth - 1); j++)
{
var aa = V.Get(x, y, (int)j);
den += aa * aa;
}
den *= this.alpha / this.n;
den += this.k;
this.S_cache_.Set(x, y, i, den); // will be useful for backprop
den = Math.Pow(den, this.beta);
A.Set(x, y, i, ai / den);
}
}
}
private void ConvOverRows(Vol V, Vol tempOutput, int inputWidth, int inputHeight, int xy_stride, int d, int ay)
{
var y = (-this.Pad | 0) + (xy_stride * ay);
var f = this.Filters[d];
for (var ax = 0; ax < this.OutputWidth; ax++) // for each out width
{
var x = (-this.Pad | 0) + (xy_stride * ax);
// convolve centered at this particular location
var a = 0.0;
for (var fy = 0; fy < f.Height; fy++) // for each element in the filter height
{
var oy = y + fy; // coordinates in the original input array coordinates
for (var fx = 0; fx < f.Width; fx++) // for each element in the filter width
{
//x is current width element of the output
//fx is the current width element of the filter
var ox = x + fx;
if (oy >= 0 && oy < inputHeight && ox >= 0 && ox < inputWidth)
{
for (var fd = 0; fd < f.Depth; fd++) // for each filter depth or input depth
{
// multiply filter pixel by image pixel and add (shared weight filter)
// avoid function call overhead (x2) for efficiency, compromise modularity :(
//filter (fx,fy,fd) *
//input (ox,oy,fd)
a += f.W[((f.Width * fy) + fx) * f.Depth + fd] *
V.W[((inputWidth * oy) + ox) * V.Depth + fd];
}
}
}
}
a += this.Biases.W[d];
tempOutput.Set(ax, ay, d, (float)a);
}
}